- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
FedCrossID studies federated person re-identification when each camera has an independent identity namespace and cross-camera identity mappings are unavailable to the normal training procedure. Protected Partial Identity Association (PPIA) uses optimal transport to match perturbed local-identity summaries, adjusts matching scores according to uncertainty, and allows identities to remain unmatched across clients. Reliability-Calibrated Relation Learning (RCRL) uses remote prototypes associated with accepted relations as supervision targets without backpropagating through those targets. Confidence weights and a round-dependent coefficient control the relation loss. Experiments on Market1501 and MSMT17 supports the utility of selected cross-client relations.
Critical strengths
-
Independent camera-local identity labels make cross-client relations difficult to use directly in local supervision. FedCrossID introduces candidate associations into training and treats relation discovery and utilization separately, providing the shared encoder with cross-camera supervision beyond local identity classification. This offers a plausible approach to label fragmentation in federated ReID.
-
PPIA performs partial matching use identity summaries and their uncertainty, with an unmatched option and mutual filtering to accommodate partially overlapping client identity sets. RCRL converts accepted associations into supervision based on remote prototypes and regulates its strength through confidence weights and a round-dependent coefficient. The two modules address which identities to associate and how to use these uncertain associations, aligning the method's design with the stated challenges.
Critical weaknesses
-
Partial optimal transport does not consistently outperform forced matching. Compared with validation-selected fixed coefficients, coefficient decay improves mAP by only 0.23 ± 0.54 and 0.26 ± 0.24 percentage points on the two datasets. Furthermore, the ablation study on the loss function involves simultaneous changes to multiple factors. Consequently, the existing evidence provides stronger support for the effectiveness of the overall approach than for the independent contributions of its individual components.
-
The authors employed "global identity mappings" to determine the association temperature. The performance of the method when parameter selection is conducted without these mappings has not been evaluated. Furthermore, the experimental conditions used for the published results cited in Table 1 differ from those employed in the authors' own implementation, making it difficult to conclude definitively that this method offers an advantage over stronger baselines.
Questions for authors
-
Is it possible to obtain retrieval results for temperatures near the current value (or for specific selected temperatures without relying on global identity mappings)? Such results would help clarify the system's sensitivity to temperature, as well as the extent to which parameter selection depends on the identity mappings established during the development phase?
-
In scenarios involving identity overlap, are there statistics available regarding the precision and recall of the adopted associations, and how do these metrics evolve over the course of training epochs? Furthermore, how do these metrics shift in response to changes in retrieval performance?
-
Switching from the legacy loss function to RCRL alters the aggregation of positive targets, the sample weighting scheme, and the normalization process simultaneously. Is there a control setup that allows for replacing confidence-based weights with uniform weights while keeping all other configurations constant? If not, what conclusions regarding confidence-based weighting can be drawn from the existing results, and which contributions remain inseparable?